Research
Building Graspable raises questions we cannot answer by opinion, so we measure. This page is where the write-ups will be listed.
Write-ups
None yet. We have not published a research write-up, and we will not list one here until it is finished and its numbers can be checked. Each will say what was measured, how, and what it does not show.
What we are working on
These are questions, not results.
How do you know an agent-built scene works?
A build that passes can still be an empty screen. We are comparing checks (building, loading in a real browser, replaying a recorded headset session) by which broken scenes each one catches and which it misses.
How far does an emulated headset go?
Testing without a headset is quick, but it is not the device. We want to state plainly which problems an emulated session finds and which only show up on real hardware.
When does a small, fast model help an agent?
Many decisions in a run are small: is this command risky, is this note worth keeping. We are measuring where a fast model can make them reliably and where it should not be trusted.
What does an agent need to remember?
Carrying context from one run to the next costs tokens and can mislead. We are measuring what is worth keeping and how to pick it for a new request.
In the meantime
The guides explain how WebXR works, with code and links to the specifications. The engineering articles describe how Graspable is built. New articles, research included, appear in the RSS feed.